chrimerss/inunda-portfolio

Inunda Flood Simulation Portfolio

GPU-accelerated 2D shallow-water flood simulations produced by Inunda, validated against USGS streamgages and satellite/SAR flood maps. Each case is a CF-compliant Zarr v3 store (time x y x x) containing simulated water depth (h, m), x-flux (u, m^2/s), and y-flux (v, m^2/s) at hourly output intervals. Grid resolution is per-case (10 m NED/FABDEM for focused US basins up to 30–90 m FABDEM/HydroSHEDS for large global domains).

🌊 Explore every case interactively in the Inunda Flood Portfolio Space → — animated 3D flood depth with a play/scrub timeline, clickable gage hydrographs (simulated vs observed water-surface rise), and satellite / SAR hit-miss-false-alarm validation overlays.

Cases

The portfolio currently spans 46 published cases across pluvial, fluvial, flash, and coastal floods worldwide. Two are detailed below as examples; the full set is listed under Case index, and every case is browsable in the interactive Space.

Case	Event	Domain	Resolution	Duration	Grid cells	Zarr size
brays_harvey	Hurricane Harvey, Aug 2017	Brays Bayou, Houston TX (HUC-12 120401040402)	10 m	6 days (2017-08-25 to 08-31)	~1.2 M	1.6 GB
guadalupe_hunt_2025	Guadalupe River flood, Jul 2025	Guadalupe above Hunt, TX (custom watershed, 744 km^2)	10 m	2 days (2025-07-03 to 07-05)	~7.4 M	344 MB

Validation summary

Metrics are on event water-surface rise (m): observed USGS gage stage-rise vs modeled depth-rise, each above its own pre-event baseline. rise-NSE isolates the rising limb (the flood response); peak dt is the peak-timing offset. Per-case metrics live in each case's gage_validation.csv (and are visualized per-gage in the Space); two representative cases:

Case	Config	Headline gage	NSE	rise-NSE	peak dt
brays_harvey	NLCD spatial Manning, constant infiltration	Brays Bayou @ MLK	0.66	0.59	0.25 h
guadalupe_hunt_2025	Uniform n=0.04, CREST LSM (cal2: wm x0.80, b x1.3, im x1.5, ksat x0.5)	N Fork Guadalupe	0.67	0.98	0.5 h

Case index

All 46 cases (each a folder at the repo root with a <case>.zarr store — see Folder structure):

• Hurricanes / major US events: brays_harvey, harvey, guadalupe_hunt_2025, florence, red_river_2019, usa
• Flash-flood benchmark (FF_USA gage-basins; NOAA MRMS + NED 10 m): ff_2021_09_NJ_ep001_01395000, ff_2021_09_NJ_ep002_01397000, ff_2021_09_PA_ep003_01451650, ff_2021_09_PA_ep003_01480500, ff_2021_09_PA_ep003_01480617, ff_2022_07_VA_ep008_03207800, ff_2023_07_VT_ep007_04293000, ff_2023_08_CA_ep003_10259100, ff_2024_06_NM_ep001_08387000, ff_2024_06_NM_ep002_08387000, ff_2024_07_NM_ep001_08387000, ff_2024_07_NM_ep005_08387000, ff_2024_09_VA_ep005_03474000, ff_2024_09_VA_ep008_03471500, ff_2024_09_VA_ep008_03473000, ff_2025_05_MD_ep001_01597500, ff_2025_05_PA_ep003_03079000, ff_2025_06_TX_ep002_08185000, ff_2025_07_NM_ep001_08387000, ff_2025_07_NM_ep002_08387000, ff_2025_07_TX_ep002_08165300, ff_2025_07_TX_ep002_08165500, ff_2025_07_TX_ep002_08166200, ff_2025_07_TX_ep002_08167000, ff_2025_07_TX_ep005_08148500
• Global remote-sensing-validated (Sen1Floods11 / NASA-IMPACT / OPERA + GEOGLOWS or IMERG): india, nigeria, spain, sri_lanka, yongchuan_2026
• UFO PlanetScope chips: ufo_bna, ufo_cmo, ufo_gil, ufo_ktm, ufo_mid, ufo_nsw, ufo_pne, ufo_que, ufo_slc, ufo_sps

Solver

• Scheme: mass-conservative local-inertial (LISFLOOD-FP; Bates et al. 2010)
• Framework: pure PyTorch, fused with torch.compile (TorchInductor)
• Time stepping: adaptive CFL (max_courant=0.4)
• Rainfall: NOAA MRMS hourly QPE (US) or GPM IMERG (global), crosswalk-gathered onto the DEM grid
• Land surface: CREST variable-infiltration-curve model or constant infiltration (per case)

Zarr structure

Each .zarr store contains:

Variable	Shape	Dtype	Description
h	(T, H, W)	float16	Water depth (m)
u	(T, H, W)	float16	x-direction unit discharge (m^2/s)
v	(T, H, W)	float16	y-direction unit discharge (m^2/s)
time	(T,)	datetime64	Output timestamps (UTC)
x	(W,)	float64	Easting coordinates (projected CRS)
y	(H,)	float64	Northing coordinates (projected CRS)
spatial_ref	scalar	—	CRS metadata (WKT)

Folder structure

Each case has its own subfolder, containing the zarr store and any additional data (e.g. gage time series):

brays_harvey/
  brays_harvey.zarr/        # depth (h), flux (u, v), coords, CRS
  gage_validation.csv       # per-gage validation metrics (NSE, rise-NSE, peak dt)
  timeseries.csv            # consolidated model + observed gage time series
  usgs/                     # raw USGS instantaneous-value data
    iv.csv
    sites.csv
guadalupe_hunt_2025/
  guadalupe_hunt_2025.zarr/ # depth (h), flux (u, v), coords, CRS
  gage_validation.csv       # per-gage validation metrics
  timeseries.csv            # consolidated model + observed gage time series
  usgs/                     # raw USGS instantaneous-value data
    iv.csv
    sites.csv

timeseries.csv columns

Column	Description
datetime_utc	Timestamp in UTC (ISO 8601)
site_no	USGS site number
site_name	USGS site name
source	observed (USGS gage) or modeled (Inunda depth at gage cell)
value	Measurement value
unit	ft (gage height), cfs (discharge), or m (modeled depth)
obs_param	USGS parameter code (00065 = gage height, 00060 = discharge) or depth for modeled

Observed records are at 15-minute intervals (native USGS IV); modeled values are at the simulation output interval (typically 1 hour), sampled as the max depth over a 5×5 cell window centered on the gage location.

Loading

import xarray as xr

# From a local clone
ds = xr.open_zarr("brays_harvey/brays_harvey.zarr")

# Or directly from Hugging Face (requires fsspec + aiohttp)
ds = xr.open_zarr(
    "hf://datasets/chrimerss/inunda-portfolio/brays_harvey/brays_harvey.zarr",
    storage_options={"anon": True},
)

# Plot max depth
ds.h.max("time").plot(cmap="Blues", vmax=3)

Citation

If you use these simulation outputs, please cite the Inunda repository:

@software{inunda2025,
  author = {Zhi, Li},
  title  = {Inunda: GPU-native flood simulator},
  year   = {2025},
  url    = {https://github.com/chrimerss/Inunda},
}

License

MIT